Bearing-TFUNet: A Triple-Fusion Segmentation Model for Bearing Defect Detection

Bearing, as a critical component in industrial production, directly dictates equipment performance. Bearing surface defects are typically minute in scale, indistinct at their boundaries, and prone to interference from surface textures and light reflections, thereby posing substantial challenges to detection. This paper proposes a novel deep segmentation network, Bearing-TFUNet, which incorporates the skip connection structure of U-Net to preserve low-level detail information, introduces an enhanced feature pyramid network (UFPN) to bolster multi-scale feature representation capabilities, and integrates a lightweight LE-Transformer for efficient modeling of global contextual information. This triple-fusion mechanism enables the decoder to concurrently integrate skip connection features from the encoder, upsampled features, and multi-scale information furnished by UFPN, facilitating comprehensive feature interaction and fusion. The DWConv-based LE-Transformer attention module effectively augments the model’s capacity for collaborative modeling of local texture details and global dependencies. Experimental results demonstrate that, compared with the baseline U-Net model, Bearing-TFUNet achieves a 12.19% improvement in the Dice coefficient, a 21.94% increase in the IoU metric, and a reduction in Hausdorff Distance from 5.6 to 4.3.